Score: graded rating

Score rates content against 2–10 ordered, described levels. The difference from Choice: levels are ordered (severity, quality, intensity), so the answer is a number you can compare and combine.

Request

{
  "type": "score",
  "instructions": "How frustrated the customer appears",
  "criteria": ["Calm, just stating facts", "Frustrated but civil", "Very angry, strong language"]
}

• criteria is an ordered array: index 0 is the lowest level; the last is the highest
• 2–10 levels (official docs, verified 2026-09-22)
• Every level needs a distinguishable description — description quality drives score quality

Answer

{
  "type": "score",
  "score": 1.0,
  "legend": { "0": "Calm…", "1": "Frustrated…", "2": "Very angry…" },
  "probabilities": { "0": 0.0, "1": 1.0, "2": 0.0 },
  "confidence": 1.0
}

Field	Meaning
score	Probability-weighted mean across levels — can be fractional (1.4 = “between frustrated and angry, leaning frustrated”)
legend	Level → description echo, handy for logs and UIs
probabilities	Distribution across levels
confidence	Concentration of the distribution; spread-out shapes (e.g. 50/50) score low

Our advice

If your business logic is “act at ≥ some threshold,” use the score number directly. To be conservative between adjacent levels, also read probabilities and gate on confidence for review.

Typical uses

• Content moderation tiers: score UGC by risk level, then publish/review/remove in code — see Content classification
• Composite scoring: the official pattern — split a complex judgment into atomic Scores and merge them with weights you control in code (via the llms.txt index). Weights are business logic; they shouldn’t live in the model
• Entity alignment: the official cookbook decides 450 candidate pairs across two beer catalogs with one three-level Score (merge / leave unlinked / hand to curator) — three levels matching three actions, with no threshold to fit

Choosing between types

• Candidates are unordered (teams, sources, skill names) → Choice
• Candidates are ordered and you want “degree” → Score
• You need a single yes/no → Noul

Next: Noul: yes-probability.